Flame AI 🔥

Flame AI is a CLI for interacting with OpenAI and doing codemods and codegen. It includes an interactive mode and several other specialized commands.

To run Flame, you need Node v18+ installed and have access to an OpenAI gpt-4 token key. If you need access to gpt-4, here's how to get it.

Flame works via npx:

npx flame --help

Commands

To see all commands, including some exciting experimental features, check out the Command Reference.

Upgrade: React Native

Flame AI's most effective command is for upgrading React Native apps.

Run this in the root of your React Native app:

# one-shot upgrade
npx flame upgrade react-native

# recommended interactive mode (much cooler):
npx flame upgrade react-native --interactive

Here's the command with all available options:

npx flame upgrade react-native --interactive --from=0.72.2 --to=auto --list --only=somefile.mm

Read more about flame upgrade react-native in the Command Reference.

AI (specifically large language models like OpenAI's ChatGPT) is a powerful tool. Flame AI is built to be a minimalist CLI that brings a great developer experience and useful tools for making code modifications (codemods) and code generation.

Given that Infinite Red is a React Native consultancy, the bulk of the built-in tools for Flame will be tuned for React Native development; however, the CLI is built to be extensible and we welcome contributions from the community.

A note about costs

Note that Flame AI costs a nominal amount of money to run, via OpenAI's API pricing. In our experiments, running a full React Native upgrade will cost about $0.05 USD or less. We frequently hit around $5 per day total cost when testing it over and over during a full work day. You must ensure you watch the CLI while it's working to ensure it doesn't get stuck in a loop and run up your bill. We are not responsible for any costs incurred by using Flame AI. We don't want those bills either, so we are regularly adding features and checks to avoid these sorts of issues. At this point, running Flame AI in an automated / unmonitored fashion is not recommended.

Limitations

You must have gpt-4 access for Flame AI to be useful. Our experiments with gpt-3.5-turbo have been very underwhelming.
The biggest limitation for FlameAI is the lack of prompt size. Currently, we're working with gpt-4, which has a token size of 8192 tokens (around 24k characters). This sounds like a lot, but considering a single code file of 200 lines can take up north of 10k characters, it is very limiting. When we have access to gpt-4-32k, we'll be able to do much more. Despite that, we've made impressive progress with the current token size.
LLMs are not perfect. They can hallucinate, miss stuff, and be generally weird. We're working on ways to mitigate this, but it's a limitation of current generation AI technology.